How Far Can a U-Net Go? An Empirical Analysis of Music Source Separation Performance

Music source separation (MSS) focuses on decomposing a mixed audio signal into individual instrumental components and is increasingly relevant for music production, restoration, remixing, education, and music information retrieval. Deep learning methods, particularly U-Net architectures operating on time–frequency representations, have recently advanced the state of the art beyond traditional signal-processing techniques. This work presents an optimized multi-source U-Net model for separating selected musical instruments from stereo mixtures. The system uses magnitude spectrograms generated by the short-time Fourier transform and is trained and evaluated on the MUSDB18 dataset. We systematically examine architectural and training-related factors, including normalization strategies, dropout placement, optimizer selection, loss weighting, data augmentation, and spectrogram-domain modifications. Separation quality is measured using BSS Eval metrics, assessing artifacts, interference, and distortion. Experimental results show that the proposed configuration achieves competitive performance relative to established convolutional and U-Net-based open-source systems, especially in terms of vocal track separation, offering practical insights into designing efficient models for multi-instrument separation.